OpenAI expands its model family, but the wait for GPT-5 continues
While enthusiasts await the launch of GPT-5, OpenAI announced a new lineup of open weight models, gpt-oss-120b and gpt-oss-20b, focusing on reasoning.
Though many have been anticipating the arrival of GPT-5, OpenAI has taken the time to expand its model offering with two open weight options that will cost less for developers to operate.
The gpt-oss-120b and gpt-oss-20b models are open-weight reasoning models the brand said are more advanced in performance than similar sized open models and can execute tool use functions and is optimized to work at the consumer hardware level. Benchmarks for gpt-oss-120b are comparable to OpenAI’s o4-mini reasoning model when running on a single 80GB GPU. Meanwhile, gpt-oss-20b compares to the o3-mini model running on edge devices with as low as 16GB of memory.
The latter is ideal for various on-device use cases, and both models have several tool offerings, including chain-of-thought (CoT) reasoning. They both support Responses API, which support agentic workflows, among other features.
A focus on safety
OpenAI highlighted the safety protocol it has deployed, particularly for its open models. The brand has also established for the gpt-oss-120b model, a special evaluation guided under its Preparedness Framework, which tests the model in similar fashions to mad actions. OpenAI also had its methodology reviewed by three independent expert groups for recommendations on how to improve its training process, and said it employed several of the suggestions.
Additionally, the brand is hosting a Red Teaming Challenge, which will include a $500K prize fund. The challenge will feature judges from OpenAI, Anthropic, Google DeepMind, and will allow participants to assist in building a safer open weight ecosystem.
Availability
The gpt-oss-120b and gpt-oss-20b models are now available under the Apache 2.0 license, and are available to developers on platforms including Azure, Hugging Face, vLLM, Ollama, llama.cpp, LM Studio, AWS, Fireworks, Together AI, Baseten, Databricks, Vercel, Cloudflare, and OpenRouter. The brand will work with hardware partners including NVIDIA, AMD, Cerebras, and Groq to support the new models. OpenAI said it will will partner with Hugging Face, Nvidia, Ollama and vLLM to host a virtual hackathon on Devpost in the coming six weeks.
The coundown to GPT-5 continues
The gpt-oss models launch not long after enthusiasts were expecting OpenAI's next foundational model, GPT-5 to arrive. While several leaks, and hints from OpenAI CEO, Sam Altman led many to believe that GPT-5 was close to launch, the state of the upcoming AI remains up in the air. Additionally, OpenAI announcing a new model ahead of GPT-5 could call into question the speculation that the AI will unify into a singular model. However, open weight models and proprietary models tend to function on different avenues.
OpenAI announced its gpt-oss models after stating earlier in the year it would make efforts to introduce more open platform products into the market, following the launch of DeepSeek’s R1 reasoning model in January. Altman stated in a Reddit AMA in February that OpenAI was “on the wrong side of history” in terms of its open-source strategy. The brand has older open-source products, such as its Whisper API, a speech-to-text model that transcribes source languages into English. However, the gpt-oss models are its first major open launch in several years.